What happens to warning capacity in AI-dependent information ecosystems?

This explores what happens to *warning capacity* — the social ability to raise an alarm and have it land — once AI saturates how knowledge is made and passed around. The corpus has a surprisingly direct answer at the source: warning isn't a content problem, it's a social act, and AI can't perform it. Can language models actually raise alarm about threats? argues that raising alarm requires three things a language model structurally lacks — felt concern, the ability to grab someone's attention rather than wait to be addressed, and the proactive initiative to speak unprompted. Models are reactive by design, and alignment training actively suppresses the kind of overclaiming a genuine warning needs. So the first thing that happens to warning capacity is that the loudest new participant in the information ecosystem simply can't sound one.

But the more interesting erosion is on the *receiving* end. Even a true warning has to be heard against background noise, and AI is flooding the channel. Can AI generate knowledge faster than humans can evaluate it? describes generation outpacing the human capacity to evaluate it — like monetary hyperinflation, every individual claim loses value because there are too many to verify, and the verification tools are themselves AI-generated, so the system can't catch its breath. When all assertions are cheap and confident, the signal that should make us sit up — "this one is urgent and true" — gets priced down to nothing along with everything else.

Lateral to this is a structural point about what AI knowledge even *is*. Does AI-generated knowledge have the same structure as hearsay? frames AI output as hearsay: testimony at a remove, mutated in every retelling, with no traceable origin and nothing stable to check it against. A warning is a claim that demands to be acted on *because* it can be trusted — but hearsay is precisely the category our verification machinery (citation, archiving, evidentiary chains) was built to discount. Why does AI output change with every prompt and context? sharpens this: AI outputs are essentially mutable, shifting with prompt and audience, so there's no fixed version of a warning to anchor accountability to. The thing that makes an alarm credible — that it points to something solid and won't change when you turn your back — is exactly what these systems lack.

Then there's the social-proof machinery that decides which signals spread at all, and it's being gamed. Why do AI posts get likes without inviting conversation? shows AI posts winning visibility through comprehensive, confident phrasing while suppressing the reply dynamics — the counter-argument, the pushback — that historically separated a legitimate alarm from mere noise. Recognition gets decoupled from validation. And in multi-agent settings the failure compounds: Why do multi-agent systems fail to coordinate at scale? finds agents accept neighbors' information without verification, propagating errors precisely because nobody raises a flag. Warning capacity isn't just muted; the network actively launders unverified claims as it scales.

The thing you didn't know you wanted to know: the deepest threat isn't that AI lies, it's that warning is *embodied and interpersonal* — it needs a someone who is concerned, addressing a someone who can be moved. Is AI returning knowledge to flow-based economies? frames AI as a return to flow-based knowledge without the embodied carrier — the speaker, the giver — that historically anchored circulation. Strip out the embodied carrier and you don't just lose a feature; you lose the channel through which alarm travels. An AI-dependent information ecosystem doesn't suppress warnings so much as dissolve the social ground that made warning possible at all.